Fast imaging of time-domain airborne EM data using deep learning technology

Abstract

Due to the huge amount of data generated by time-domain airborne electromagnetic (AEM) systems, conductivity depth imaging methods are widely used to help in the interpretation of these data because they can be generated quickly and easily. We have introduced a new imaging method generated using a deep neural network. The network structure combines four convolutional neural networks with a long short-term memory technique and adopts an error back-propagation scheme to update the parameters. A deep hierarchical structure is used to extract and store the complex nonlinear relationship between the model and electromagnetic (EM) responses, creating results close to 1D inversions. To check the effectiveness, we have examined our algorithm on synthetic and survey AEM data. The imaging result shows that this method, when using reasonable network parameters, can not only image well at high speed, but the method also is not very sensitive to noise. Another advantage of this method is that once the training is completed on a well-distributed training set, the network can be used without any changes to process other data sets from an identically configured AEM system.